| --- |
| license: apache-2.0 |
| library_name: pytorch |
| tags: |
| - ocr |
| - text-detection |
| - text-recognition |
| - easyocr |
| - craft |
| - safetensors |
| --- |
| |
| # EasyOCR weights (CRAFT detector and English recogniser) in safetensors |
|
|
| The two pretrained models the [EasyOCR](https://github.com/JaidedAI/EasyOCR) `en` reader uses, converted from the |
| official `.pth` files to `.safetensors`. Nothing was retrained or modified: every tensor is identical, in value and dtype, |
| to the original, and the key names (including the `module.` prefix) are unchanged. |
|
|
| | File | Model | Tensors | SHA-256 | |
| |---|---|---|---| |
| | `craft_mlt_25k.safetensors` | CRAFT text detector | 154 | `4ab7f69bd8161b437f18f1718b8f6e58136bd081639f238b567e987415c20268` | |
| | `english_g2.safetensors` | English recogniser (CRNN, "generation 2") | 44 | `307a47690148280d00e4314b1868ebce44089d0974903a56fe7822902d1b31d0` | |
|
|
| `SHA256SUMS` lists the same values: `sha256sum -c SHA256SUMS`. |
|
|
| ## Origin |
|
|
| Converted from the official EasyOCR release assets (EasyOCR's own MD5 values in `easyocr/config.py` are shown): |
|
|
| | Original | Release asset | MD5 of the original `.pth` | |
| |---|---|---| |
| | `craft_mlt_25k.pth` | https://github.com/JaidedAI/EasyOCR/releases/download/pre-v1.1.6/craft_mlt_25k.zip | `2f8227d2def4037cdb3b34389dcf9ec1` | |
| | `english_g2.pth` | https://github.com/JaidedAI/EasyOCR/releases/download/v1.3/english_g2.zip | `5864788e1821be9e454ec108d61b887d` | |
| |
| Conversion: `torch.load(..., weights_only=True)` of the state dict, then `safetensors.torch.save_file`. The saved file was |
| loaded back and compared tensor by tensor with the original (all equal). An EasyOCR reader fed these tensors returned the |
| same boxes, text and confidences as the original reader on the test images it was checked with (CPU). |
| |
| ## Use |
| |
| These are state dicts, not a ready `easyocr.Reader` folder: EasyOCR itself only looks for MD5-checked `.pth` files, so the |
| detector and recogniser have to be built from the state dicts by the caller. |
| |
| ```python |
| from safetensors.torch import load_file |
|
|
| craft = load_file("craft_mlt_25k.safetensors") # CRAFT state dict (keys start with "module.") |
| recog = load_file("english_g2.safetensors") # recogniser state dict (keys start with "module.") |
| ``` |
| |
| ## Licence and attribution |
| |
| The files are redistributed under the licences of the projects they come from. Keep these notices with them. |
| |
| - **EasyOCR**, JaidedAI: Apache License 2.0. https://github.com/JaidedAI/EasyOCR |
| - **CRAFT** (`craft_mlt_25k`), Clova AI Research, NAVER Corp.: MIT License. https://github.com/clovaai/CRAFT-pytorch. |
| Paper: Baek et al., "Character Region Awareness for Text Detection", CVPR 2019 (https://arxiv.org/abs/1904.01941). |
| EasyOCR ships Clova AI's pretrained CRAFT model. |
| - **Recognition training pipeline**, a modified version of `clovaai/deep-text-recognition-benchmark`: Apache License 2.0. |
| https://github.com/clovaai/deep-text-recognition-benchmark. Model architecture: CRNN (Shi et al., 2015). |
| |
| The upstream projects publish licences for their code. The weight files carry no separate licence statement that was found, |
| so they are assumed to follow the repositories above. This repository is an unofficial conversion and is not affiliated |
| with JaidedAI or Clova AI. |
| |
